Cropping intensity and double-crop detection from time series

Dense vegetation-index time series reveal how many crop cycles a field completes each year, information that static imagery cannot provide. Harmonic regression on Sentinel-2, Landsat 8/9 and MODIS stacks separates single, double and triple-cropped land with enough confidence to feed national food-balance models.

Sensors

Sentinel-2 MSI (ESA): 10 m spatial resolution in visible and near-infrared bands; 5-day revisit at the equator with both satellites operating. The red-edge bands (B5, B6, B7 at 20 m) are particularly sensitive to early green-up and late senescence, extending the usable phenological window beyond what broadband NDVI alone captures.
Landsat 8/9 OLI (USGS/NASA): 30 m resolution with a 16-day revisit per satellite; combined Landsat 8 and 9 operations reduce this to roughly 8 days. The archive extends to 1984 for Landsat 5, giving multi-decadal trend context. Surface-reflectance Collection 2 products are the standard input for phenological time-series work.
MODIS MOD13Q1 (NASA Terra): 250 m NDVI and EVI composites at 16-day intervals. Coarse resolution makes field-level mapping unreliable below roughly 1 ha, but the dense, cloud-composited time series is well-suited to regional cropping-intensity mapping and calibration of harmonic models before applying them to finer-resolution stacks.
Sentinel-2 and Landsat fusion stacks: Combining both constellations through spatial-temporal adaptive reflectance fusion (STARFM and variants) can produce synthetic daily or near-daily surface-reflectance imagery at 10–30 m. This matters most in humid tropical regions where either sensor alone rarely achieves cloud-free coverage within a single crop cycle.

What the time series actually measures

A vegetation-index time series is a record of canopy greenness sampled repeatedly across the growing season. Each crop cycle produces a characteristic arc: a rapid green-up phase as leaf area expands, a plateau at peak biomass, then a decline through senescence and harvest. A double-cropped field produces two such arcs within twelve months. A triple-cropped field, common in parts of the Mekong Delta and the Indo-Gangetic Plain, produces three.

The key indices are NDVI (the ratio of near-infrared to red reflectance) and EVI (which adds a blue-band correction for aerosol and canopy background effects). EVI is generally preferred in high-biomass tropical settings because it saturates less aggressively. Neither index is a direct measure of yield or crop type; they measure greenness, which is a proxy for leaf area and chlorophyll content. That distinction matters when interpreting outputs.

Harmonic regression: fitting cycles to noise

Raw time series are messy. Cloud gaps, aerosol contamination and bidirectional reflectance effects all introduce noise. Harmonic regression, sometimes called Fourier-based phenology fitting, addresses this by fitting sinusoidal components of known frequencies to the observed data. A single annual harmonic captures one crop cycle; adding a semi-annual harmonic captures two. The amplitude and phase of each fitted harmonic carry agronomic meaning: amplitude correlates with peak greenness intensity, phase gives the timing of green-up and senescence.

The TIMESAT software package, developed at Lund University and widely used in peer-reviewed phenology studies, formalises this approach and allows extraction of phenological metrics including start of season, end of season, length of growing period and number of growing seasons per year. Inflection-point methods, which look for local minima and maxima in smoothed time series rather than fitting global functions, are an alternative that handles asymmetric crop cycles better in some settings. Both approaches are in routine operational use.

Cloud cover is the dominant practical constraint. In the humid tropics, a single Sentinel-2 or Landsat overpass may be obscured for weeks at a time. Fusion approaches and MODIS gap-filling can partially compensate, but the honest position is that cropping-intensity maps in persistently cloudy regions carry higher uncertainty than those in semi-arid irrigated zones where clear-sky observations are frequent.

Resolution floors and what they mean for smallholders

MODIS at 250 m is adequate for national-scale cropping-intensity mapping but cannot resolve fields smaller than roughly one hectare without severe mixed-pixel contamination. Sentinel-2 at 10 m resolves fields down to about 0.1 ha under favourable conditions, which covers most smallholder plots in South and Southeast Asia. Landsat at 30 m sits between these extremes.

The practical implication is that programme design should match sensor choice to field-size distribution. A food-balance model for a country dominated by large irrigated schemes can be built on MODIS with Landsat validation. A programme serving smallholder-dominated landscapes, such as the fragmented plots common across sub-Saharan Africa, needs Sentinel-2 as the primary layer. Even then, fields below roughly 0.2 ha remain difficult to classify reliably because adjacent land covers contaminate the pixel signal.

What irrigated intensity reveals that area statistics miss

National agricultural statistics typically report harvested area, not cropping intensity. A country that double-crops 40 percent of its irrigated land produces substantially more food than its planted-area figure suggests, and vice versa. Satellite-derived cropping-intensity maps close this gap by providing an independent, spatially explicit count of harvest events.

The practical users of these maps fall into two groups. Food-balance modellers use intensity layers to adjust production estimates derived from yield models: a pixel classified as double-cropped contributes twice to the annual production sum. Irrigation planners use them differently, to identify where intensification is already occurring, where water demand is therefore highest, and where groundwater drawdown risk may be concentrated. Neither application requires the map to be perfect; it needs to be consistent across years so that trends are detectable. Interannual consistency is something well-calibrated harmonic models handle reasonably well, provided the input archive is homogeneous.

Accuracy, confusion and honest caveats

Published validation studies on MODIS-based global cropping-intensity maps report overall accuracies in the range of 80 to 90 percent for distinguishing single from double cropping in major agricultural regions, with lower accuracy in transitional zones where farmers shift intensity year to year in response to water availability or market prices. Sentinel-2-based studies in specific regions report higher figures, sometimes above 90 percent, but these are typically validated against ground-truth data collected in the same season, which is an optimistic test.

The most common error is confusing a long-fallow period mid-season with a harvest event, or misclassifying a perennial crop with a seasonal flush as a double-cropped annual. Perennial tree crops are a known source of false positives. Careful masking of non-cropland areas using an independent land-cover layer reduces but does not eliminate this problem. Users should treat intensity maps as probabilistic classifications, not census data.

Satellize applied phenological time-series methods in its crop-estimation programme for the Kingdom of Tonga, where the small island context required careful handling of mixed pixels along field boundaries and coastal margins. The same methodological principles apply at continental scale, though the cloud and archive challenges differ substantially.

Archive depth and what historical stacks make possible

The Landsat archive back to 1984 and the MODIS archive from 2000 make it possible to reconstruct decadal trends in cropping intensity. This matters for questions such as whether irrigated double-cropping has expanded in response to groundwater access, whether climate shifts have altered the length of the second growing season, or whether intensification collapsed during a drought year. These are questions that field surveys cannot answer retrospectively.

Sentinel-2 data begins in 2015 for Sentinel-2A and 2017 for Sentinel-2B, giving a shorter but higher-resolution archive. For most operational programmes, the practical approach is to use MODIS for trend context back to 2000, Landsat for the intermediate period at 30 m, and Sentinel-2 for current-season mapping at 10 m. Outputs from the three sensors are not directly comparable without careful cross-calibration, and any report that presents a continuous 25-year Sentinel-equivalent time series without explaining that step should be read with scepticism.

Typical figures

Primary spatial resolution	10 m (Sentinel-2 MSI), 30 m (Landsat 8/9 OLI), 250 m (MODIS MOD13Q1)
Effective revisit (cloud-free)	5 days at equator (Sentinel-2A+B combined); ~8 days (Landsat 8+9 combined); 16-day composites (MODIS MOD13Q1)
Minimum mappable field size	~0.1 ha (Sentinel-2); ~0.5 ha (Landsat); ~1 ha (MODIS, indicative)
Spectral bands used	Red (660 nm), NIR (835 nm), red-edge (705–783 nm, Sentinel-2 only), blue (490 nm, for EVI)
Archive depth	Landsat from 1984; MODIS from 2000; Sentinel-2 from 2015
Primary vegetation indices	NDVI, EVI (MODIS standard product); NDVI, EVI, red-edge NDVI (Sentinel-2 derived)
Typical mapping latency	End-of-season maps: 2–4 weeks post-harvest; near-real-time intensity updates: 10–16 days depending on compositing window
Cropping-intensity classes	Single, double, triple crop per calendar year; fallow; non-cropland (masked)
Typical overall accuracy (published range)	80–90% for single vs double distinction in major irrigated systems; lower in transitional or mixed-crop zones
Delivery formats	GeoTIFF raster (per-pixel class and confidence), vector polygon summary, CSV statistics by administrative unit

Analytics Satellize can run

Annual cropping-intensity map	Harmonic regression (Fourier decomposition) on annual NDVI/EVI stack; semi-annual harmonic amplitude threshold for double-crop classification	GeoTIFF with per-pixel intensity class (single/double/triple/fallow) and amplitude confidence layer
Phenological metrics extraction	TIMESAT-style inflection-point detection on smoothed Sentinel-2 or Landsat time series; start, peak and end of each growing season per pixel	Multi-band GeoTIFF of phenological dates (day-of-year) per season, delivered as GIS layer
Interannual intensity trend report	Mann-Kendall trend test applied to annual cropping-intensity fraction over MODIS or Landsat archive (2000–present); pixel-level and zonal aggregation	PDF trend report with mapped trend significance and tabular summary by district or irrigation command area
Harvested-area adjustment factor for food-balance models	Cropping-intensity map multiplied against planted-area layer to produce effective harvested-area estimate; uncertainty bounds from per-pixel confidence scores	CSV table of adjusted harvested area by crop season and administrative unit, with uncertainty range
In-season double-crop detection alert	Running harmonic fit updated on each new Sentinel-2 acquisition; alert triggered when second green-up inflection detected above amplitude threshold	Automated GIS feed or email alert with mapped extent of confirmed second-cycle fields, updated at each cloud-free overpass
Fusion-enhanced time series for cloudy regions	STARFM spatial-temporal fusion of Sentinel-2 and MODIS to generate synthetic 10 m daily surface-reflectance estimates; harmonic fitting applied to fused stack	Dense time-series raster stack and derived intensity map, with data-gap log indicating fusion-reliant dates

Sources

Who does the work

We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.